Agentic RAG is retrieval-augmented generation in which the retrieval itself is driven by an autonomous agent loop rather than a single search. The system plans sub-queries, executes them across multiple retrieval tools, observes what came back, reflects on what is still missing, and refines — before synthesising one final, cited answer.
The loop: plan → act → observe → reflect
- Plan. The question is decomposed: “Compare the termination clauses of our three largest suppliers” becomes sub-queries per supplier plus a comparison step.
- Act. Sub-queries run in parallel against multiple tools: vector search, knowledge-graph neighbourhood expansion, community summaries.
- Observe & reflect. A critic model checks the evidence for gaps and contradictions and issues refined follow-up queries where needed.
- Synthesise. The final answer is generated from the de-duplicated evidence pool, with numbered citations back to the source documents.
Why bounded loops matter in the enterprise
An unbounded agent is a liability: unpredictable latency, unpredictable cost, unauditable behaviour. Production agentic RAG needs hard guardrails. In SovraRAG the loop is capped at a fixed number of iterations and a hard timeout, evidence per step is capped, every step is logged to the query log, and any failure falls back to classic single-pass retrieval — so the answer path is always auditable and the worst case is always defined.
When agentic RAG wins
Simple lookups do not need an agent — and a good router should not use one. Agentic retrieval pays off on comparative, aggregating and multi-hop questions: portfolio-wide clause comparisons, exposure mapping, “what changed between policy generations”, literature reviews. In SovraRAG the router detects question complexity automatically; in internal multi-hop evaluations the agentic path beat single-pass retrieval on every quality metric while being faster at the median, because sub-queries run concurrently.
Agentic RAG and data sovereignty
An agent that fans out dozens of retrievals amplifies your exposure if those calls leave your jurisdiction. SovraRAG's agentic loop runs entirely on Swiss infrastructure — language models, embeddings and all retrieval infrastructure — so multiplying retrieval steps never multiplies cross-border risk.